Friday, August 12, 2008

Mathematical Statistics Preliminary Examination

Statistics Group, Department of Mathematics and Statistics, Auburn University

Name:

1. It is a closed-book and in-class exam.

2. One page (letter size, 8.5-by-11in) cheat sheet is allowed.

3. Calculator is allowed. No laptop (or equivalent).

4. Show your work to receive full credits. Highlight your final answer.

5. Solve any five problems out of the seven problems.

6. Total points are 50. Each question is worth 10 points.

7. If you work out more than five problems, your score is the sum of five highest points.

8. Time: 180 minutes. (8:30am - 11:30am, Friday, August 12, 2011)

1 2 3 4 5 6 7 Total

1
1. Suppose that X1 , X2 , and X3 are mutually independent uniform(0, 1). Calculate

P (max{X1 , X2 } X3 ).

You need to calculate it rigorously using density functions to get full points.

2. Let U1 , . . . , Un be a random sample from the uniform(0, 1) distribution. Let U(s) be

the sth order statistic. Determine E(U(s) ) and V ar(U(s) ).

3. Suppose a population consists of the integers 1, . . . , N . Suppose we draw a sample

using binomial sampling in the following manner: we flip a fair coin N times (once for
each of integer in the population) and the integer is included if the coin lands heads.
Let W be the sum of the integers in the sample. Compute E(W ) and V ar(W ).

4. Let tn be a sequence of real-valued estimators such that for R, the expectation

and variance of tn exist and E(tn ) and V ar(tn ) 0 as n . Show that tn is
consistent for .

5. Suppose that X1 , . . . , Xn are iid with density

3! x2 "
f (x, ) = 1 2 , |x| and > 0.
4

(a) Find an estimate using the method of moments.

(b) Find the asymptotic distribution of as n . (Hint: use the Central Limit
Theorem and delta method)

6. Suppose that X1 , . . . , Xn are iid N (, 2 ) random variables. Find the uniformly min-
imum variance unbiased estimator (UMVUE) of 2 / 2 . Hint: for any function h and
a chi-squared random variable 2d , we have E[h(2p )] = p E[h(2p+2 )/2p+2 ].

7. Suppose that X1 , . . . , Xn are iid N (, 2 ) and Y1 , . . . , Ym are iid N (, 2 ). Derive the

likelihood ratio test for

H0 : = 2 + 1, Ha : '= 2 + 1

Note that the two samples have different means but the same variance. To get full
points, you should try your best to simplify the rejection region.

2
 Solutions

1. The probability is

P (max{X1 , X2 } X3 )
= P (X1 X3 ) + P (X1 < X3 , X2 X3 )
= E[P (X1 X3 |X3 )] + E[P (X1 < X3 |X3 )P (X2 X3 |X3 )]

For X uniform(0, 1) and 0 t 1, we have

P (X < t) = t, P (X t) = 1 t

Therefore,

P (max{X1 , X2 } X3 ) = E(1 X3 ) + E[X3 (1 X3 )]

2
= 1 E(X32 ) = .
3

Alternative approach: We may argue that with equal probability Xi is the largest
number. Since uniform(0, 1) is a continuous distribution, we do not need to consider
the case where some of Xi are equal. Therefore,
2
P (max{X1 , X2 } X3 ) = P (X1 is largest) + P (X2 is largest) = .
3
This second solution will not get full points.

2. Recall that the pdf of the sth order statistic for a random sample from a distribution
F with density f is
# $
n1
n [F (x)]s1 [1 F (x)]ns f (x) .
s1

For the uniform(0, 1) distribution, this becomes

# $
n 1 s1
n x (1 x)ns , 0<x<1.
s1

This is the density of the beta distribution B(s, n s + 1). Thus

s s(n s + 1)
E(Us ) = and V ar(U(s) ) = .
n+1 (n + 1)2 (n + 2)

3. The sum of the values in the sample is

W = 1I1 + 2I2 + + N IN ,

3
 where Ij is 1 or 0 as j is included in the sample for j = 1, . . . , N . But
1
E(Ij ) = P (Ij = 1) =
2
and
1
V ar(Ij ) = E(Ij2 ) E 2 (Ij ) = E(Ij ) E 2 (Ij ) = .
4
Thus
N
1% N (N + 1)
E(W ) = j=
2 j=1 4
and, since I1 , . . . , IN are independent,
N
1 % 2 N (N + 1)(2N + 1)
V ar(W ) = j = .
4 j=1 24

4. By Chebychevs inequality, for any $ > 0, we have

E({tn }2 ) V (tn ) + {E(tn ) }2
0 P (|tn | $) = .
$2 $2
Since E(tn ) and V (tn ) 0 as n , we have P (|tn | $) 0 as n .
5. (a) The expectation is
&
E(X) = xf (x)dx = 0

The second moment is

1 2
&
2
E(X ) = x2 f (x)dx =
5
Therefore, we can match the the second moment with its sample counterpart, and
'
( n
(5 %
= ) x2
n i=1 i

(b) According to the Central Limit Theorem, we have

! 1 % 2 "
n n xi N (0, 2 ), as n

where = E(X 2 ) and 2 = var(X 2 ). Noticing that = 5, we have

n( ) N (0, 2 ( 5/(2 ))2 ), as n .
We can calculate 2 as follows,
3 4 1 8 4
2 = E(X 4 ) E(X 2 )2 = 4 =
35 25 175
Therefore,

n( ) N (0, (2/7)2 ), as n .

4
6. We know that the sample mean X N (, 2 /n) and the sample variance S 2 have
(n 1)S 2 / 2 2n1 . Additionally X is independent of S 2 . It is also known that
(x, s2 ) is a sufficient and complete statistic. Try to calculate the expectation of X 2 /S 2 .
n1
E(X 2 /S 2 ) = E(X 2 )E(1/S 2 ) = (2 + 2 /n) E( )
2 n1
n1 1 n 1 2 1
= (2 + 2 /n) = ( + )
2 n 3 n 3 2 n
Therefore,
n 3 X 2 1 2
E( ) =
n 1 S2 n 2
n3 X 2
The UMVUE of 2 / 2 is n1 S 2
n1 .

7. The log-likelihood function is

n m
n 1 % m 1 %
%(, , 2 ) = log(2 2 ) 2 (xi )2 log(2 2 ) 2 (yi )2
2 2 i=1 2 2 i=1

The unrestricted MLEs can be solved from

n
% 1 %
= 2 (xi ) = 0
i=1
m
% 1 %
= 2 (yi ) = 0
i=1
n m
% n 1 % 2 m 1 %
= 2 + 4 (xi ) 2 + 4 (yi )2 = 0
2 2 2 i=1 2 2 i=1

Therefore, the unrestricted MLEs are

*n
x)2 + m 2
*
2 i=1 (xi i=1 (yi y)
= x, = y, =
m+n

If we let = 2 + 1, then the likelihood becomes

n m
2 n 2 1 % 2 m 2 1 %
%(, ) = log(2 ) 2 (xi 2 1) log(2 ) 2 (yi )2
2 2 i=1 2 2 i=1

The restricted MLEs can be solved from

n m
% 1 % 1 %
= 2 2(xi 2 1) + 2 (yi ) = 0
i=1 i=1
n m
% n 1 % 2 m 1 %
= 2 + 4 (xi 2 1) 2 + 4 (yi )2
2 2 2 i=1 2 2 i=1

5
Therefore, the restricted MLEs are
*n
x)2 + m 2
*
2nx + my 2n 2 i=1 (xi i=1 (yi y)
= , = 2 + 1, =
4n + m m+n
The likelihood ratio test statistic is

%(, , 2 )
= = ( 2 / 2 )m+n
%(, , 2 )

Therefore, the rejection region is

2 / 2 < c

where c is a constant in (0, 1) and satisfies P ( 2 / 2 < c) = if H0 is true.